“I was very surprised. I didn’t expect to lose.” That was the moment 9th Dan professional Go player Lee Sedol realized just how far the field of Artificial Intelligence has advanced. It was also the moment he realized that the $1Million up for grabs could be lining the virtual pocket of his opponent, AlphaGo, rather than his.
The words were uttered following his game one defeat to a program developed by Google’s DeepMind unit in a historic series of the ancient Chinese strategy game, Go in March 2016. Just understand that Go, unlike other games such as Chess where computers dispose of world champions with embarrassing ease, has long been considered a challenge for AI. Why? Because of the intuition required. That and the sheer number of possible legal positions on the 19×19 square grid on which it’s played which, when written in full, looks like this:
208 168 199 381 979 984 699 478 633 344 862 770 286 522 453 884 530 548 425 639 456 820 927 419 612 738 015 378 525 648 451 698 519 643 907 259 916 015 628 128 546 089 888 314 427 129 715 319 317 557 736 620 397 247 064 840 935
That’s a lot of permutations for a human to program in order to “teach” a machine. It doesn’t exactly lend itself to an “assess all options” kind of strategy. Rather, it needs to narrow down those options and think much more like a human does. It needs to truly learn the game. Therein lies the significance of this result: this Go victory represents an Artificial Intelligence milestone in deep learning capability. And that milestone represents endless potential for innovation — potential we want to see fulfilled.
Bragging rights? Not quite
The team behind the victor, DeepMind, was a London-based AI start-up bought by Google in a deal reportedly worth around $500Million. Now, either Google paid some serious money just for the bragging rights of outsmarting Mr Sedol, or they recognized the importance of AI in the future of innovation. Given the price tag, probably the latter. And they’re not alone. There’s a long list of young, innovative AI companies that have been acquired by big players including Google, IBM, Yahoo, Intel, Apple and Salesforce since 2011 and nearly 60% of them had VC backing. Indeed such investment reflects the current state of the art: we are currently undergoing an AI boom.
Such a boom is reflected not only in the internet world but in the academic world too. More and more AI labs are being set up under the control of renowned scientists and academics who impart their knowledge to others via online courses for example. The result: inaccessibility to AI knowledge is not holding ambitious startups back from developing boundary-pushing applications.
Diving in the deep end
But what is causing this boom? Why the sudden acceleration in the last 5 or so years? To answer these, it is necessary to first dig deeper: so let’s get down and technical.
First; some jargon-busting. “Neural networks” is a term you’ve probably heard thrown around. It isn’t new. It sits at the back of the AI research cupboard and is a type of computer architecture on which AI is built. Structurally, it resembles the human brain — nodes connected in a web that tackle problems collectively — hence the name. On this neural network you might run a “machine learning” program whereby you teach machines to reach a solution of their own by showing them vast numbers of cases until they learn what to do. “Deep learning” is the next layer and it is this that the big players are interested in. It is this that opens up boundless possibilities for innovation.
Deep learning is a type of machine learning that uses complex algorithms across multiple layers of neural networks — each one analyzing data at a different level of abstraction, building up to a program that can eventually understand even the most complicated data — like face recognition (believe me, that’s up there in terms of difficulty for machines). However when training machines, it is necessary to subject them to extremely high volumes of training sets — and for deep learning, this requires computing power, troves of data and time.
Thankfully, and this is where we get back to explaining the boom, in the last few years, computing speed has continued to evolve exponentially and the use of GPUs (which can perform mathematical operations simultaneously) has continued to increase. Also, thanks to the internet, we aren’t exactly short of user-generated data and research into more efficient algorithms is in full swing — with predictive or real-time learning algorithms poised to be the next big thing. On top of that, more and more big players are open-sourcing software to open up their research. That and access to faster prototyping means the stage is set for something big.
There’s learning and then there’s application
Crawling back out of that deep technical hole and returning to our opening gambit, of course the victory in Go meant more than just a victory in Go. There are many such examples of illustrations of progress in AI but how can this progress be applied to commercial success? How can it impact global industry? That’s what we at next47 are interested in.
We see this field as opening the door to fantastic opportunities for digital industrial applications. We are eager to seize these opportunities and turn them into concrete solutions. Just think of the potential applications of physical, autonomous systems in industry alone. As was mentioned in our Autonomous Machine article, AI is the bedrock for Industry 4.0 and its collaborative, adaptive, flexible manufacturing. Whether it’s using predictive data analytics to improve production capacity; providing new insights to optimize production workflows; improving preventative (self) maintenance and repair; improving defect detection and real time ordering, the possibilities of machine learning in manufacturing are endless. And it’s not only the building of products but the moving of them too. Our partnering startup Magazino is currently developing and building perception-controlled, mobile robots for intralogistics. Thanks to AI, they have the ability to identify individual objects on a shelf, securely grasp it and finally place it precisely at their destination.
Like the machines going through deep learning, we still have much to learn. We know we can only shape the future of innovation through exchange with creative partners. So let’s get together and teach these machines a thing or two.